Tech Page 2

Researchers at Caltech and Stanford have unveiled HomeBody, a system that gives GPT Astra persistent spatial memory and the ability to execute compound humanoid actions for long-horizon tasks. In a kitchen demo, a Unitree G1 robot guided by Astra cleaned the room and retrieved remembered items from ambiguous requests. The team said the system requires no environment-specific training or additional policy learning. HomeBody replaces the trained intermediate VLA layer in conventional three-part architectures—System2 VLM for vision and instructions, System1 VLA for control commands, and System0 for coordinated motion. As frontier models such as Astra improve, the researchers argue trained VLAs are no longer necessary and System2 can directly manage reusable motion skills. A plug-and-play VLM calls a composable skill library....

Boston Dynamics has introduced a new four-fingered hand for its Atlas humanoid, a 13-degree-of-freedom, directly actuated design intended for mass production. The hand marks a strategic shift. Earlier Atlas hands had seven degrees of freedom and were designed to grasp a wide range of objects. The new version focuses on manipulating them. “We considered many, many designs,” said Alberto Rodriguez, director of robot behavior for Atlas. “It boils down to a trade-off of competing objectives.” The design omits a pinky and is tailored for high-fidelity simulation and sim-to-real reinforcement learning. It uses a single actuator type, fully encapsulated joints and no fragile cables crossing joints. Boston Dynamics said the hand can slide the thumb tip along and across the other...

Figure, the U.S. humanoid robot developer, has released Helix 2.5, its most advanced neural network model to date, and said it enabled zero-shot whole-body autonomy in 30 real homes. Figure called it the first demonstration at that scale of zero-shot whole-body behaviour generalisation for a humanoid robot. The company said a single pre-trained foundation model was applied to three types of behaviour: locomotion; manipulation of rigid and deformable objects; and bimanual coordination, as well as active perception. Figure said that, with task-specific data, architecture, training method and evaluation unchanged, pre-training on Index alone raised zero-shot recognition success to 56% from 9%. Index is Figure’s robot dataset project built for global scale. The model also validated a Human-to-Humanoid Robot Transfer Scaling...

What happens if robot model training data is expanded from 300 hours to 30,000 hours? Can robotic systems, similar to large language models, continuously gain new capabilities as datasets grow? AGIBOT’s newly released native World Action Model (WAM), GE-Act 2.0, systematically answers that question for the first time. Unlike mainstream approaches that build upon existing video generation models, every parameter of GE-Act 2.0 — including core modules for visual representation, future prediction and action forecasting — is initialized from scratch and trained end-to-end on embodied manipulation data. After pre-training, the model undergoes zero-shot real robot tests on unfamiliar scenes and unseen objects without fine-tuning or demonstration for benchmark tasks. The evaluation covers 100 atomic tasks, 20 skill categories and two...

Chinese robotics company Unitree said it has fully open-sourced an upgraded version of its general-purpose humanoid foundation model, UnifoLM-WLA-1.0, setting new state-of-the-art results on multiple global open-source embodied model benchmarks. The model can perform both tabletop manipulation and whole-body mobile manipulation with a single model, enabling cross-task and cross-end-effector generalization, Unitree said. It achieves a “one-model-driven, whole-body coordination” effect, the company added. Unitree said UnifoLM-WLA-1.0 was trained on large-scale general-purpose multimodal perception and understanding datasets and incorporates interaction-centric world modeling, significantly strengthening spatial perception and scene understanding. It outperformed several open-source competitors in China and overseas on multiple embodied reasoning benchmarks, with some capabilities comparable to mainstream closed-source models, according to the company. In real-robot tests, a single model completed...

ACE ROBOTICS, in collaboration with the Chinese University of Hong Kong’s Multimedia Laboratory and Nanyang Technological University, has open-sourced ACE-Ego-Hand, a 3D dual-hand shape reconstruction model designed for first-person video, generating approximately 5,000 hours of training data that could serve as a foundational dataset for embodied AI model training. The model extracts structured 3D bimanual manipulation trajectories from massive egocentric video footage, translating human grasping, assembly and hand-coordination actions into machine-learnable data. It marks the first time a video generation model has been repurposed as a DreamHand encoder for end-to-end training, outputting hand pose, shape and 3D position in a single forward pass without frame-by-frame detection or temporal regression. Efficiency gains over traditional approaches Unlike previous hand models reliant on...

Tingbo He, head of semiconductor research at Huawei Technologies, published a new academic paper on Thursday addressing concerns about heat dissipation in 3D-stacked chips, marking the company’s third paper on its “τ-law” chip design philosophy in less than four months. The latest paper, titled “Huawei’s τ Chip Was Supposed to Melt?”, responds to scepticism that folding circuits into three dimensions would inevitably lead to higher power density and thermal issues. “Heat is the most troublesome hidden concern of the τ-law,” She wrote. “Inside the chip, the main energy consumption does not come from computation itself, but from data transfer.” Contrary to expectations that circuit folding would dramatically increase transistor density and power density, the paper argues that folding actually reduces...

Chinese robotics firms LEJU Robot and Heyman have jointly launched a scene-oriented solution centred on a general-purpose humanoid robot body and a multi-fingered dexterous hand, aiming to eliminate the need for custom end-effector swaps across diverse industrial tasks. The system addresses common pain points such as trade-offs between payload and precision, low gesture-development efficiency, and difficult redeployment, the companies said. The integrated dexterous hand features a variable-configuration design that supports seven plus grip modes – including two-finger pinch, internal expansion, enveloping grasp, and two-handed hold – on the same hardware. It delivers a payload of at least 24 kg and a repeated positioning accuracy of ≤0.03 mm, outperforming the industry norm of 0.2 mm. This allows a single manipulator to...

Lyte, a U.S.-based physical AI perception infrastructure startup, has raised $165 million in a Series C funding round, valuing the company at $1.6 billion post-money, the company said. The round was led by Maverick Silicon, with participation from existing investors including Fidelity Management & Research, Atreides Management, Key1 Capital, and Ora Global. Proceeds will be used to boost sensor chip production, scale its LyteVision integrated perception platform, and expand its workforce. Founded in 2021, Lyte emerged from stealth in January this year and had previously raised approximately $107 million. The company develops chips, multimodal sensors, and spatial perception software to help robots navigate environments more safely. Its flagship unified spatial intelligence platform, LyteGalaxy, integrates sensors, compute, software, and algorithms into...

Astribot, a Chinese robotics startup, has unveiled SmoothRL, an online reinforcement learning framework designed to train policies during asynchronous execution – a growing reality for embodied AI deployments where inference and action run at different frequencies. The work, led by the company’s foundation model team, addresses a fundamental mismatch: when a robot receives new action chunks before previous ones finish, conventional RL that updates on entire chunks ends up crediting or blaming actions that were never actually executed. The framework partitions each action chunk into three regions – committed (already fixed), execution (actually performed), and discarded (overwritten by later inference) – and backpropagates gradients only through the execution region. This ensures that optimisation mirrors the robot’s real experience. SmoothRL also...